[docs] static model rules

[stevhliu]

adds docs for static model rules so model contributors are aware of expectations

I wonder if it wouldn't make sense to auto generate that portion of the doc automatically

added from @tarekziade feedback:

• reformats format_rule_details() to output Markdown that gets inserted directly in the docs
• adds a check_modeling_rules_doc.py util that checks rules.toml and uses format_rule_details() to generate the ## Rules reference section
• adds util to make check-repo and make fix-repo to catch any drifts and automatically regenerate the docs if a rule is added or removed from rules.toml

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[tarekziade]

nice!
when we your run mlinter --rule you get a documentation snippet, very similar to your doc for each rule. See also --list-rules.

I wonder if it wouldn't make sense to auto generate that portion of the doc automatically

[stevhliu]

oh good idea!

i think it'd be clearer and easier to scan if we combined the bad_example and good_example fields with a single diff field for each rule in rules.toml. wdyt?

[tarekziade]

oh good idea!

i think it'd be clearer and easier to scan if we combined the bad_example and good_example fields with a single diff field for each rule in rules.toml. wdyt?

Could you show me an example of what you have in mind? would that be a reorg in the toml or in how the CLI outputs it?

[stevhliu]

i was thinking a reorg in rules.toml to something like:

[rules.TRF001.explanation]
what_it_does = "Checks naming consistency between <Model>PreTrainedModel and config_class."
why_bad = "Mismatched config_class can break loading, auto classes, and developer expectations."
diff = '''
   class AcmePreTrainedModel(PreTrainedModel):
-  config_class = NanoConfig
+  config_class = AcmeConfig
'''

[tarekziade]

i was thinking a reorg in rules.toml to something like:

[rules.TRF001.explanation]
what_it_does = "Checks naming consistency between <Model>PreTrainedModel and config_class."
why_bad = "Mismatched config_class can break loading, auto classes, and developer expectations."
diff = '''
   class AcmePreTrainedModel(PreTrainedModel):
-  config_class = NanoConfig
+  config_class = AcmeConfig
'''

nice! +1 to do this.

[tarekziade]

super nice addition, thanks!

[stevhliu]

do you have permissions to merge for me @tarekziade? i don't think the required failing test is related here

[tarekziade]

I don't have the perms. cc @ydshieh